from richard miller: /n/sources/patch/applied/rpi-model-a-usb
Empirical adjustments to usb for Raspberry Pi Model A (which has no built-in hub).

Reference: /n/atom/patch/applied2013/rpia
Date: Mon Jun  3 05:50:03 CES 2013
Signed-off-by: quanstro@quanstro.net

--- /sys/src/9/bcm/usbdwc.c	Mon Jun  3 05:48:57 2013
+++ /sys/src/9/bcm/usbdwc.c	Mon Jun  3 05:48:57 2013
@@ -24,38 +24,37 @@
 
 #include "dwcotg.h"
 
-#define	USBREGS	(VIRTIO+0x980000)
-
 enum
 {
-	Enabledelay		= 50,
-	Resetdelay		= 10,
+	USBREGS		= VIRTIO + 0x980000,
+	Enabledelay	= 50,
+	Resetdelay	= 10,
 	ResetdelayHS	= 50,
 
-	Read			= 0,
-	Write			= 1,
+	Read		= 0,
+	Write		= 1,
 };
 
 typedef struct Ctlr Ctlr;
 typedef struct Epio Epio;
 
 struct Ctlr {
-	Dwcregs		*regs;		/* controller registers */
-	int			nchan;		/* number of host channels */
-	ulong		chanbusy;	/* bitmap of in-use channels */
-	QLock		chanlock;	/* serialise access to chanbusy */
-	QLock		split;		/* serialise split transactions */
-	int			splitretry;	/* count retries of Nyet */
-	int			sofchan;	/* bitmap of channels waiting for sof */
-	int			wakechan;	/* bitmap of channels to wakeup after fiq */
-	int			debugchan;	/* bitmap of channels for interrupt debug */
-	Rendez		*chanintr;	/* sleep till interrupt on channel N */
+	Dwcregs	*regs;		/* controller registers */
+	int	nchan;		/* number of host channels */
+	ulong	chanbusy;	/* bitmap of in-use channels */
+	QLock	chanlock;	/* serialise access to chanbusy */
+	QLock	split;		/* serialise split transactions */
+	int	splitretry;	/* count retries of Nyet */
+	int	sofchan;	/* bitmap of channels waiting for sof */
+	int	wakechan;	/* bitmap of channels to wakeup after fiq */
+	int	debugchan;	/* bitmap of channels for interrupt debug */
+	Rendez	*chanintr;	/* sleep till interrupt on channel N */
 };
 
 struct Epio {
 	QLock;
-	Block		*cb;
-	ulong		lastpoll;
+	Block	*cb;
+	ulong	lastpoll;
 };
 
 static Ctlr dwc;
@@ -77,8 +76,8 @@
 	qlock(&ctlr->chanlock);
 	bitmap = ctlr->chanbusy;
 	for(i = 0; i < ctlr->nchan; i++)
-		if((bitmap&(1<<i)) == 0){
-			ctlr->chanbusy = bitmap|(1<<i);
+		if((bitmap & (1<<i)) == 0){
+			ctlr->chanbusy = bitmap | 1<<i;
 			qunlock(&ctlr->chanlock);
 			return &ctlr->regs->hchan[i];
 		}
@@ -107,9 +106,9 @@
 	Ctlr *ctlr = ep->hp->aux;
 
 	if(ep->debug)
-		ctlr->debugchan |= 1<<(hc - ctlr->regs->hchan);
+		ctlr->debugchan |= 1 << (hc - ctlr->regs->hchan);
 	else
-		ctlr->debugchan &= ~(1<<(hc - ctlr->regs->hchan));
+		ctlr->debugchan &= ~(1 << (hc - ctlr->regs->hchan));
 	switch(ep->dev->state){
 	case Dconfig:
 	case Dreset:
@@ -117,6 +116,7 @@
 		break;
 	default:
 		hcc = ep->dev->nb<<ODevaddr;
+		break;
 	}
 	hcc |= ep->maxpkt | 1<<OMulticnt | ep->nb<<OEpnum;
 	switch(ep->ttype){
@@ -138,10 +138,15 @@
 		hcc |= Lspddev;
 		/* fall through */
 	case Fullspeed:
-		hc->hcsplt = Spltena | POS_ALL | ep->dev->hub<<OHubaddr | ep->dev->port;
-		break;
+		if(ep->dev->hub > 1){
+			hc->hcsplt = Spltena | POS_ALL | ep->dev->hub<<OHubaddr |
+				ep->dev->port;
+			break;
+		}
+		/* fall through */
 	default:
 		hc->hcsplt = 0;
+		break;
 	}
 	hc->hcchar = hcc;
 	hc->hcint = ~0;
@@ -153,7 +158,7 @@
 	Dwcregs *r;
 
 	r = a;
-	return (r->gintsts & Sofintr);
+	return r->gintsts & Sofintr;
 }
 
 static void
@@ -170,7 +175,7 @@
 		r->gintmsk |= Sofintr;
 		sleep(&ctlr->chanintr[n], sofdone, r);
 		splx(x);
-	}while((r->hfnum&7) == 6);
+	}while((r->hfnum & 7) == 6);
 }
 
 static int
@@ -179,52 +184,62 @@
 	Hostchan *hc;
 
 	hc = a;
+	if(hc->hcint == (Chhltd|Ack))
+		return 0;
 	return (hc->hcint & hc->hcintmsk) != 0;
 }
 
 static int
 chanwait(Ep *ep, Ctlr *ctlr, Hostchan *hc, int mask)
 {
+	int intr, n, x, ointr;
 	ulong start, now;
 	Dwcregs *r;
-	int intr, n, x;
-	int ointr;
 
 	r = ctlr->regs;
 	n = hc - r->hchan;
 	for(;;){
-	restart:
+restart:
 		x = splfhi();
 		r->haintmsk |= 1<<n;
 		hc->hcintmsk = mask;
 		sleep(&ctlr->chanintr[n], chandone, hc);
 		hc->hcintmsk = 0;
 		splx(x);
-		if(((intr = hc->hcint)&Chhltd) != 0)
+		intr = hc->hcint;
+		if(intr & Chhltd)
 			return intr;
 		start = fastticks(0);
 		ointr = intr;
 		now = start;
 		do{
-			if(((intr = hc->hcint)&Chhltd) != 0){
+			intr = hc->hcint;
+			if(intr & Chhltd){
 				if((ointr != Ack && ointr != (Ack|Xfercomp)) ||
-				   intr != (Ack|Chhltd|Xfercomp) || (now - start) > 60)
-					dprint("await %x after %ld %x -> %x\n", mask, now - start, ointr, intr);
+				   intr != (Ack|Chhltd|Xfercomp) ||
+				   (now - start) > 60)
+					dprint("await %x after %ld %x -> %x\n",
+						mask, now - start, ointr, intr);
 				return intr;
 			}
-			if((intr&mask) == 0){
-				dprint("ep%d.%d await %x intr %x -> %x\n", ep->dev->nb, ep->nb, mask, ointr, intr);
+			if((intr & mask) == 0){
+				dprint("ep%d.%d await %x intr %x -> %x\n",
+					ep->dev->nb, ep->nb, mask, ointr, intr);
 				goto restart;
 			}
 			now = fastticks(0);
 		}while(now - start < 100);
-		dprint("ep%d.%d halting channel %8.8ux hcchar %8.8ux grxstsr %8.8ux gnptxsts %8.8ux hptxsts %8.8ux\n", ep->dev->nb, ep->nb, intr, hc->hcchar, r->grxstsr, r->gnptxsts, r->hptxsts);
+		dprint("ep%d.%d halting channel %8.8ux hcchar %8.8ux "
+			"grxstsr %8.8ux gnptxsts %8.8ux hptxsts %8.8ux\n",
+			ep->dev->nb, ep->nb, intr, hc->hcchar, r->grxstsr,
+			r->gnptxsts, r->hptxsts);
 		mask = Chhltd;
 		hc->hcchar |= Chdis;
 		start = m->ticks;
-		while(hc->hcchar&Chen){
+		while(hc->hcchar & Chen){
 			if(m->ticks - start >= 100){
-				print("ep%d.%d channel won't halt hcchar %8.8ux\n", ep->dev->nb, ep->nb, hc->hcchar);
+				print("ep%d.%d channel won't halt hcchar %8.8ux\n",
+					ep->dev->nb, ep->nb, hc->hcchar);
 				break;
 			}
 		}
@@ -252,7 +267,7 @@
 			return 0;
 	}else
 		return 0;
-	if(hc->hcchar&Chen){
+	if(hc->hcchar & Chen){
 		iprint("hcchar %8.8ux hcint %8.8ux", hc->hcchar, hc->hcint);
 		hc->hcchar |= Chen | Chdis;
 		while(hc->hcchar&Chen)
@@ -306,7 +321,8 @@
 		return;
 	p = chanlog[0];
 	for(i = 0; i < nchanlog; i++){
-		print("%5.5d.%5.5d %8.8ux %8.8ux %8.8ux %8.8ux\n", p[0]&0xFFFF, p[0]>>16, p[1], p[2], p[3], p[4]);
+		print("%5.5d.%5.5d %8.8ux %8.8ux %8.8ux %8.8ux\n",
+			p[0]&0xFFFF, p[0]>>16, p[1], p[2], p[3], p[4]);
 		p += 5;
 	}
 	nchanlog = 0;
@@ -340,17 +356,19 @@
 		hctsiz = hc->hctsiz;
 		hc->hctsiz = hctsiz & ~Dopng;
 		if(hc->hcchar&Chen){
-			dprint("ep%d.%d before chanio hcchar=%8.8ux\n", ep->dev->nb, ep->nb, hc->hcchar);
+			dprint("ep%d.%d before chanio hcchar=%8.8ux\n",
+				ep->dev->nb, ep->nb, hc->hcchar);
 			hc->hcchar |= Chen | Chdis;
 			while(hc->hcchar&Chen)
 				;
 			hc->hcint = Chhltd;
 		}
 		if((i = hc->hcint) != 0){
-			dprint("ep%d.%d before chanio hcint=%8.8ux\n", ep->dev->nb, ep->nb, i);
+			dprint("ep%d.%d before chanio hcint=%8.8ux\n",
+				ep->dev->nb, ep->nb, i);
 			hc->hcint = i;
 		}
-		if(hc->hcsplt&Spltena){
+		if(hc->hcsplt & Spltena){
 			qlock(&ctlr->split);
 			sofwait(ctlr, hc - ctlr->regs->hchan);
 			if((dwc.regs->hfnum & 1) == 0)
@@ -361,72 +379,98 @@
 		hc->hcchar = (hc->hcchar &~ Chdis) | Chen;
 		clog(ep, hc);
 		if(ep->ttype == Tbulk && dir == Epin)
-			i = chanwait(ep, ctlr, hc, /* Ack| */Chhltd);
-		else if(ep->ttype == Tintr && (hc->hcsplt&Spltena))
+			i = chanwait(ep, ctlr, hc, /* Ack| */ Chhltd);
+		else if(ep->ttype == Tintr && (hc->hcsplt & Spltena))
 			i = chanwait(ep, ctlr, hc, Chhltd);
 		else
 			i = chanwait(ep, ctlr, hc, Chhltd|Nak);
 		clog(ep, hc);
 		hc->hcint = i;
 
-		if(hc->hcsplt&Spltena){
+		if(hc->hcsplt & Spltena){
 			hc->hcsplt &= ~Compsplt;
 			qunlock(&ctlr->split);
 		}
 
-		if((i&Xfercomp) == 0 && i != (Chhltd|Ack) && i != Chhltd){
-			if(i&Stall)
+		if((i & Xfercomp) == 0 && i != (Chhltd|Ack) && i != Chhltd){
+			if(i & Stall)
 				error(Estalled);
-			if(i&Nyet)
+			if(i & (Nyet|Frmovrun))
 				continue;
-			if(i&Nak){
+			if(i & Nak){
 				if(ep->ttype == Tintr)
 					tsleep(&up->sleep, return0, 0, ep->pollival);
 				else
 					tsleep(&up->sleep, return0, 0, 1);
 				continue;
 			}
+			logdump(ep);
 			print("usbotg: ep%d.%d error intr %8.8ux\n",
 				ep->dev->nb, ep->nb, i);
-			if(i&~(Chhltd|Ack))
+			if(i & ~(Chhltd|Ack))
 				error(Eio);
 			if(hc->hcdma != hcdma)
-				print("usbotg: weird hcdma %x->%x intr %x->%x\n", hcdma, hc->hcdma, i, hc->hcint);
+				print("usbotg: weird hcdma %x->%x intr %x->%x\n",
+					hcdma, hc->hcdma, i, hc->hcint);
 		}
 		n = hc->hcdma - hcdma;
 		if(n == 0){
-			if((hc->hctsiz&Pktcnt) != (hctsiz&Pktcnt))
+			if((hc->hctsiz & Pktcnt) != (hctsiz & Pktcnt))
 				break;
 			else
 				continue;
 		}
 		if(dir == Epin && ep->ttype == Tbulk && n == nleft){
-			nt = (hctsiz&Xfersize) - (hc->hctsiz&Xfersize);
+			nt = (hctsiz & Xfersize) - (hc->hctsiz & Xfersize);
 			if(nt != n){
-				if(n == ((nt+3)&~3))
+				if(n == ROUND(nt, 4))
 					n = nt;
 				else
-					print("usbotg: intr %8.8ux dma %8.8ux-%8.8ux hctsiz %8.8ux-%8.ux\n",
-						i, hcdma, hc->hcdma, hctsiz, hc->hctsiz);
+					print("usbotg: intr %8.8ux "
+						"dma %8.8ux-%8.8ux "
+						"hctsiz %8.8ux-%8.ux\n",
+						i, hcdma, hc->hcdma, hctsiz,
+						hc->hctsiz);
 			}
 		}
 		if(n > nleft){
-			if(n != ((nleft+3)&~3))
-				dprint("too much: wanted %d got %d\n", len, len - nleft + n);
+			if(n != ROUND(nleft, 4))
+				dprint("too much: wanted %d got %d\n",
+					len, len - nleft + n);
 			n = nleft;
 		}
-		if((nleft -= n) == 0 || (n % maxpkt) != 0)
+		nleft -= n;
+		if(nleft == 0 || (n % maxpkt) != 0)
+			break;
+		if((i & Xfercomp) && ep->ttype != Tctl)
 			break;
-		if((i&Xfercomp) && ep->ttype != Tctl)
-				break;
 		if(dir == Epout)
-			dprint("too little: nleft %d hcdma %x->%x hctsiz %x->%x intr %x\n", nleft, hcdma, hc->hcdma, hctsiz, hc->hctsiz, i);
+			dprint("too little: nleft %d hcdma %x->%x hctsiz %x->%x intr %x\n",
+				nleft, hcdma, hc->hcdma, hctsiz, hc->hctsiz, i);
 	}
 	logdump(ep);
 	return len - nleft;
 }
 
 static long
+multitrans(Ep *ep, Hostchan *hc, int rw, void *a, long n)
+{
+	long sofar, m;
+
+	sofar = 0;
+	do{
+		m = n - sofar;
+		if(m > ep->maxpkt)
+			m = ep->maxpkt;
+		m = chanio(ep, hc, rw == Read? Epin : Epout, ep->toggle[rw],
+			(char*)a + sofar, m);
+		ep->toggle[rw] = hc->hctsiz & Pid;
+		sofar += m;
+	}while(sofar < n && m == ep->maxpkt);
+	return sofar;
+}
+
+static long
 eptrans(Ep *ep, int rw, void *a, long n)
 {
 	Hostchan *hc;
@@ -447,20 +491,11 @@
 		nexterror();
 	}
 	chansetup(hc, ep);
-	if(rw == Read && ep->ttype == Tbulk){
-		long sofar, m;
-		sofar = 0;
-		do{
-			m = n - sofar;
-			if(m > ep->maxpkt)
-				m = ep->maxpkt;
-			m = chanio(ep, hc, Epin, ep->toggle[rw], (char*)a + sofar, m);
-			ep->toggle[rw] = hc->hctsiz & Pid;
-			sofar += m;
-		}while(sofar < n && m == ep->maxpkt);
-		n = sofar;
-	}else{
-		n = chanio(ep, hc, rw == Read? Epin : Epout, ep->toggle[rw], a, n);
+	if(rw == Read && ep->ttype == Tbulk)
+		n = multitrans(ep, hc, rw, a, n);
+	else{
+		n = chanio(ep, hc, rw == Read? Epin : Epout, ep->toggle[rw],
+			a, n);
 		ep->toggle[rw] = hc->hctsiz & Pid;
 	}
 	chanrelease(ep, hc);
@@ -509,7 +544,11 @@
 	chansetup(hc, ep);
 	chanio(ep, hc, Epout, SETUP, req, Rsetuplen);
 	if(req[Rtype] & Rd2h){
-		b->wp += chanio(ep, hc, Epin, DATA1, data, datalen);
+		if(ep->dev->hub <= 1){
+			ep->toggle[Read] = DATA1;
+			b->wp += multitrans(ep, hc, Read, data, datalen);
+		}else
+			b->wp += chanio(ep, hc, Epin, DATA1, data, datalen);
 		chanio(ep, hc, Epout, DATA1, nil, 0);
 		n = Rsetuplen;
 	}else{
@@ -548,7 +587,7 @@
 greset(Dwcregs *r, int bits)
 {
 	r->grstctl |= bits;
-	while(r->grstctl&bits)
+	while(r->grstctl & bits)
 		;
 	microdelay(10);
 }
@@ -563,7 +602,7 @@
 	ctlr = hp->aux;
 	r = ctlr->regs;
 
-	ctlr->nchan = 1 + ((r->ghwcfg2&Num_host_chan) >> ONum_host_chan);
+	ctlr->nchan = 1 + ((r->ghwcfg2 & Num_host_chan) >> ONum_host_chan);
 	ctlr->chanintr = malloc(ctlr->nchan * sizeof(Rendez));
 
 	r->gahbcfg = 0;
@@ -577,14 +616,14 @@
 	tsleep(&up->sleep, return0, 0, 25);
 	r->gahbcfg |= Dmaenable;
 
-	n = (r->ghwcfg3&Dfifo_depth)>>ODfifo_depth;
+	n = (r->ghwcfg3 & Dfifo_depth) >> ODfifo_depth;
 	rx = 0x306;
 	tx = 0x100;
 	ptx = 0x200;
 	r->grxfsiz = rx;
 	r->gnptxfsiz = rx | tx<<ODepth;
 	tsleep(&up->sleep, return0, 0, 1);
-	r->hptxfsiz = (rx+tx) | ptx<<ODepth;
+	r->hptxfsiz = (rx + tx) | ptx << ODepth;
 	greset(r, Rxfflsh);
 	r->grstctl = TXF_ALL;
 	greset(r, Txfflsh);
@@ -665,7 +704,8 @@
 static void
 epopen(Ep *ep)
 {
-	ddprint("usbotg: epopen ep%d.%d ttype %d\n", ep->dev->nb, ep->nb, ep->ttype);
+	ddprint("usbotg: epopen ep%d.%d ttype %d\n",
+		ep->dev->nb, ep->nb, ep->ttype);
 	switch(ep->ttype){
 	case Tnone:
 		error(Enotconfig);
@@ -677,6 +717,7 @@
 			ep->toggle[Read] = DATA0;
 		if(ep->toggle[Write] == 0)
 			ep->toggle[Write] = DATA0;
+		break;
 	}
 	ep->aux = malloc(sizeof(Epio));
 	if(ep->aux == nil)
@@ -686,13 +727,15 @@
 static void
 epclose(Ep *ep)
 {
-	ddprint("usbotg: epclose ep%d.%d ttype %d\n", ep->dev->nb, ep->nb, ep->ttype);
+	ddprint("usbotg: epclose ep%d.%d ttype %d\n",
+		ep->dev->nb, ep->nb, ep->ttype);
 	switch(ep->ttype){
 	case Tctl:
 		freeb(((Epio*)ep->aux)->cb);
 		/* fall through */
 	default:
 		free(ep->aux);
+		break;
 	}
 }
 
@@ -830,10 +873,11 @@
 	r->hport0 = Prtpwr;
 	tsleep(&up->sleep, return0, 0, Enabledelay);
 	s = r->hport0;
-	if((b = (s&(Prtconndet|Prtenchng|Prtovrcurrchng))) != 0)
+	b = s & (Prtconndet|Prtenchng|Prtovrcurrchng);
+	if(b != 0)
 		r->hport0 = Prtpwr | b;
 	dprint("usbotg reset=%d; sts %#x\n", on, s);
-	if((s&Prtena) == 0)
+	if((s & Prtena) == 0)
 		print("usbotg: host port not enabled after reset");
 	return 0;
 }
@@ -849,7 +893,8 @@
 	ctlr = hp->aux;
 	r = ctlr->regs;
 	s = r->hport0;
-	if((b = (s&(Prtconndet|Prtenchng|Prtovrcurrchng))) != 0)
+	b = s & (Prtconndet|Prtenchng|Prtovrcurrchng);
+	if(b != 0)
 		r->hport0 = Prtpwr | b;
 	b = 0;
 	if(s & Prtconnsts)